Performance profiling and analysis

Report
Accelerating TauDEM as a Scalable
Hydrological Terrain Analysis Service on XSEDE
Ye Fan1, Yan Liu1, Shaowen Wang1, David Tarboton2, Ahmet
Yildirim2, Nancy Wilkins-Diehr3
1 University
of Illinois at Urbana-Champaign
Utah State University
3 San Diego Supercomputer
2
XSEDE’14
Atlanta, GA, July 15, 2014
1
Outline
• Introduction
o TauDEM software
o Parallelism
o ECSS work plan
• Computational Intensity Analysis and Performance
Improvement
o Strategies
o Findings & results
• TauDEM Gateway Application
o Data integration
o Workflow construction
o XSEDE-enabled execution
2
Scalable DEM-based Hydrological
Information Analysis
• Digital Elevation Models (DEM)
o Geospatial topographic data
o Raster and vector representation
• DEM-based Hydrological Information Analysis
o Use of topographic information in hydrological analysis and modeling
o Examples
• Derivation of flow directions, contributing area, stream network…
• Impact of High Resolution DEM Data
o High resolution DEM data sources
• National Elevation Dataset (NED) from the U.S. Geological Survey (USGS)
o 10-meter resolution: 330GB raw data
o 1-meter resolution: 4-5 PB raw data
• OpenTopography Lidar-derived DEM data
o Improved accuracy and reliability of analysis and modeling results
o Revealing insights that were not possible to obtain before
3
Example: USGS NED
4
TauDEM
• TauDEM - A Parallel Computing Solution to DEM-based
Terrain Analysis
o Open source software
o A suite of DEM tools for the extraction and analysis of hydrologic information from
topographic data
o A growing user community
• Parallel Computing in TauDEM
o Parallel programming model: Message Passing Interface (MPI)
o Spatial data decomposition
• Each process reads a sub-region for processing
• MPI communication for exchanging runtime hydrological information
• Each process writes a sub-region defined by output data decomposition
o Parallel input/output (IO)
• In-house GeoTIFF library (no support for big GeoTIFF)
• MPI IO for DEM read and write
5
TauDEM Channel Network and Watershed Delineation Software
• Stream and watershed
delineation
• Multiple flow direction flow
field
• Calculation of flow-based
derivative surfaces
http://hydrology.usu.edu/taudem/
Multi-File Input Model
Number of processes
mpiexec –n 5 pitremove ...
results in the domain being
partitioned into 5 horizontal 5
stripes
On input files (red rectangles)
data coverage may be
arbitrarily positioned and may
overlap or not fill domain
completely. All files in the folder
are taken to comprise the
domain.
Only limit is that no one file is
larger than 4 GB.
Maximum GeoTIFF file size: 4
GB = about 32000 x 32000
rows and columns
Multi-File Output Model
Number of processes
mpiexec –n 5 pitremove ...
results in the domain being
partitioned into 5 horizontal 5
stripes
2 rows of
files per
stripe
Multifile option
-mf 3 2
results in each stripe being
output as a tiling of 3 columns
and 2 rows of files
Maximum GeoTIFF file size: 4
GB = about 32000 x 32000
rows and columns
3 columns of files per stripe
Computational Challenges
• Scalability issues
o PitRemove step on 2GB DEM
• 681 seconds on an 8-core PC
• 3,759 seconds on a 64-core cluster
• Not acceptable on XSEDE resources
• Computational challenges
o Scaling to large-scale analysis using massive computing
resources is difficult
o Cyberinfrastructure-based computational analysis needs
in-depth knowledge and expertise on computational
performance profiling and analysis
9
Computational Scaling Issues
• Results collected on local cluster with Network File System (NFS)
interconnet
• YellowStone dataset (27814x19320)
o
Using more processors reduced compute time, but suffered from longer execution time
• Chesapeake dataset (53248x53248)
o
Execution could not finish on D8FlowDir operation
http://hydrology.usu.edu/taudem/taudem5.0/TauDEM_4_12.pptx
10
CyberGIS-OT-TauDEM Collaboration
TauDEM 5.0
Lidarderived
DEMs
OpenTopography
OT TauDEM
Services
Scalability
Enhancement
(XSEDE ECSS)
TauDEM 5.x
CyberGIS
DEMs
USGS NED
OT
User DEMs
CyberGISTauDEM App
TauDEM-enabled
Research
11
ECSS Goals
• Enhance TauDEM for large-scale
terrain analysis on massive
computing resources provided on
national cyberinfrastructure
through rigorous computational
performance profiling and analysis
12
Collaboration Team
• National cyberinfrastructure
o
o
Extreme Science and Engineering Discovery Environment (XSEDE)
XSEDE Extended Collaborative Support Services (ECSS) provides computational science expertise
• Ye Fan, Yan Liu, Shaowen Wang, National Center for Supercomputing Applications (NCSA)
• NSF OpenTopography LiDAR data facility
o
o
o
DEM generation services for LiDAR-derived TauDEM analysis
Integration of TauDEM in OpenTopography service environment
People
• Chaitan Baru, Nancy Wilkins-Diehr, Choonhan Yeon, San Diego Supercomputer Center
(SDSC)
• NSF CyberGIS project
o
o
o
Integration of TauDEM in CyberGIS Gateway
Integration of TauDEM in advanced CyberGIS analytical services (workflow)
People
• University of Illinois at Urbana-Champaign (UIUC)
o Yan Liu, Anand Padmanabhan, Shaowen Wang
• San Diego Supercomputer Center (SDSC)
o Nancy Wilkins-Diehr, Choonhan Yeon
13
Performance Analysis: Challenges
• System-level performance variation is very difficult to identify
o Computing seemed not the reason for performance slowdown
o Network issue or file system issue? NFS is difficult to debug
• Barrier for performance profiling
o Performance profiling tools deployment need system
administration skills
o Using performance profiling libraries may need code change
o Configuring profiling parameters and interpreting profiling
results are not trivial
14
Strategies
•
•
•
Project management
o Code repository
• TauDEM source code is moved to github to facilitate multi-party development
and testing
• http://github.com/dtarb/TauDEM
o Documentation
• Github wiki
• Google Drive
o Meetings
• Bi-weekly teleconference
Build and test
o XSEDE resources:
• [email protected]: for tests using up to 1,024 processors
• [email protected]: for tests using up to 16,384 processors
o Profiling tools
• Darshan: I/O profiling
Performance profiling and analysis
o Computational bottleneck analysis
• Focus on I/O performance
o Scalability to processors
o Scalability to data size
o Performance optimization
15
Generic I/O Profiling
• Darshan profiling
found anomaly on file
read operations
• The finding is
confirmed in TauDEM
timing data
16
IO Bottlenecks - Input
• Inefficient File Reading
o n processes, m files
o Original version: n x m file reads for getting geo-metadata
o Fix: 1 x m file reads + MPI_Bcast
• Coding Issue
o File read deadlock situation caused by too many opened file descriptors
o File not closed in time
o Fix: close a file as soon as read operation is done
17
IO Bottleneck - Output
• Inefficient MPI IO
o Original spatial domain decomposition did not consider IO performance
o Improvement: domain decomposition strategy is changed to reduce the number of
processes needed by an output file
• No Collective IO
• Parallel File System
o Use as many OSTs on Lustre file system
18
Scalability Results
• Scalability Tests
o Processors: up to 1,024
o Data sizes: 2GB, 12GB, 36GB DEMs
• IO No Longer a Bottleneck
19
Results – Resolving I/O Bottlenecks
#cores
Compute
Header Read Data Read
Data Write
32
42.7 / 42.8
193.5 / 3.8
0.4 / 0.4
153.5 / 3.5
64
35.3 / 34.8
605.5 / 3.9
1.5 / 1.1
160.2 / 2.3
128
33.7 / 33.0
615.2 / 2.6
0.9 / 1.0
173.2 / 2.3
256
37.5 / 38.0
831.7 / 2.3
0.5 / 0.9
391.3 / 1.6
Table 1. I/O Time Comparison (before / after; in seconds)
(Fan et al. 2014)
20
Results – Execution Time
Figure 2. Execution time of the three most costly TauDEM functions on a
36GB DEM dataset. (Fan et al. 2014)
21
Next Steps
• More Room to Improve
o41.6 hours using 1024 cores on 36GB
DEM
• Communication Pattern Analysis
• Methodological Investigation on Algorithm
Design
22
CyberGIS-TauDEM Gateway Application
• Streamlined TauDEM Analysis in CyberGIS Gateway
o
o
o
o
Web environment
Transparent integration of DEM data sources
Customized TauDEM analysis workflow
Online visualization
• Status
o 2 prototypes in April and May, respectively
o Alpha release in early July
o Beta release in August
23
CyberGIS Application Integration Framework
GISolve Middleware
Controller
DB
Data
Selection
Analysis
Input
Panels
Geo-Input
Editing
Job
Job Wrappers
Data
Visualization
Workflow
Mapping
Visualization
Sharing
Data Storage
Computing
Environment
CyberGIS Gateway
Job Panel
CyberGIS Toolkit
Data Retrieval
Geo Data
Processing
Execution Setup
Data
Servers
Metadata
Servers
Parallel
Computing
Post-processing
Mapping
Servers
External
Data
Sources
Geo-visualization
24
Data Integration
• Multiple High Resolution DEM Sources
o USGS NED (10-meter)
• Hosted at UIUC
• Map preview
o OpenTopography LiDAR-derived DEMs
• Web service API
• Data Retrieval
o USGS NED: wget
o OT: Dynamic DEM generation and downloading
o Data caching
• XWFS?
• Data Processing
o
o
o
o
o
Study area clipping
Multi-file generation
Reprojection
GDAL library (http://gdal.org)
High-performance map reprojection
• Collaborative work with USGS
25
Analysis Workflow
• Approach
o 26 TauDEM functions
o Template-based customization of TauDEM functions
• Pre-defined dependency
• Dynamic workflow construction in Gateway
• Data format: JSON
• Implementation
o Interactive workflow configuration
• Ext JS + SigmaJS
• Execution
o Runtime command sequence generation
• On Trestles: command sequence
• On Stampede: a set of jobs linked based on job dependency
26
Visualization
• Visualization Computation
o Reprojection
o Pyramid generation for multiple zoom levels
o Coloring (symbology)
• Online Visualization
o Each product is a map layer accessible through the OGC-standard Web Mapping
Service (WMS)
27
DEMO
28
Concluding Discussions
•
•
•
•
Multidisciplinary collaboration is a key to the success so far
Great potential for further performance improvement
Performance profiling and analysis at large scale is critical
Guidelines for future software research and development
o Explicit computational thinking in software development lifecycle (design, coding, testing)
o Performance analysis remains challenging.
o Collaboration with computational scientists and conducting performance profiling on
cyberinfrastructure are important
o Cyberinfrastructure provides a set of abundant and diverse computational platforms for
identifying computational bottlenecks and scaling code performance
• CyberGIS-TauDEM Gateway application significantly lowers the
barrier of conducting large-scale TauDEM analyses by community
users
29
Acknowledgements
• XSEDE (NSF Grant No. 1053575)
• This material is based in part upon work supported by
NSF under Grant Numbers 0846655 and 1047916
• TauDEM team work is supported by the US Army
Research and Development Center contract No.
W912HZ-11-P-0338 and the NSF Grant Number
1135482.
• Any opinions, findings, and conclusions or
recommendations expressed in this material are those of
the author and do not necessarily reflect the views of the
National Science Foundation
30
Thanks!
31

similar documents